Forests play a vital role in the wellbeing of our planet. Large- and small-scale deforestation across the globe is threatening the stability of our climate, forest biodiversity, and thus the preservation of fragile ecosystems and our natural habitat as a whole. With increasing public interest in climate change issues and forest preservation, a large demand for carbon offsetting, carbon footprint ratings, and environmental impact assessments is emerging. Satellite remote sensing is the only method that can provide global coverage at frequent revisit times and is therefore the standard method for global forest monitoring. Most often, deforestation maps are created from optical data such as Landsat and MODIS. Although such maps are of generally good quality, they cannot quantify biomass and are not typically available at less than annual intervals due to persistent cloud cover in many parts of the world, especially the tropics where most of the world's forest biomass is concentrated. Synthetic Aperture Radar (SAR) can fill this gap as it penetrates clouds and interacts with the three-dimensional structure on the ground in a way that scales with volume and therefore biomass. While longer wavelengths are better for deeper penetration of the canopy and therefore biomass estimation, one of the most readily available data sources is Sentinel-1, a shorter wavelength C-band SAR. We have previously shown that Sentinel-1 data can achieve good separability of forest and non-forest globally (Hansen et al. 2020).

A big challenge for developing and validating algorithms for deforestation detection is the scarcity of reliable reference data. Ideally, the exact time and type of change is known for a sufficiently large area and time period to be used for training and validation. However, data of this kind are infeasible to obtain because they require continuous large-scale monitoring and are therefore prohibitively expensive. For global coverage, the best available forest maps are still only updated at annual intervals, e.g. GlobalForestWatch (Hansen et al. 2013) or JRC Tropical Moist Forests (Vancutsem et al. 2021) and are themselves obtained from Machine Learning models. In order to circumvent the reliance on reference data to train deforestation detection algorithms, one could simply apply change detection algorithms that do not rely on training labels such as Bayesian Online Changepoint Detection (Adams & MacKay 2007) or PELT (Killick et al. 2012). However, this approach has the downside that the detected changes do not necessarily correspond to deforestation. Detected changes may instead reflect land cover transitions other than forest to non-forest, or indeed measurement changes that do not represent a change in the underlying land cover at all, for example:

- Seasonal changes in vegetation
- Growth and harvest cycles in agriculture
- Soil moisture changes due to rainfall

This is because such change detection algorithms are in no way specific to any particular type of change but merely pick up statistically significant changes in the raw data, whether or not these changes correspond to a change in the underlying state. To solve these problems, and to mitigate the disadvantages of both the fully supervised and the fully unsupervised methods, we have pursued a partially supervised approach: instead of requiring reference data that capture land cover changes and thus need a temporal component, our method only relies on a stationary forest map and thus classifies pixels as either being stable forest or not. Pixels that are not stable forest could include agriculture, urban areas, other forms of vegetation, etc., or pixels that undergo deforestation at some point. This reference forest map can then be used to detect pixels that deviate from the reference class, i.e., the forest prototype. This is done by computing a distance metric between the time series of the pixel and an ensemble of reference forest time series. Deforestation can then be detected as a sudden increase in distance to the reference forest class when computed over time. Preliminary results show that this method achieves a near-perfect change detection sensitivity (producer's accuracy above 99%), although false positives occasionally lead to a low user's accuracy of about 60%. The mean change detection delay amounts to about two to three months. Further work is expected to reduce the false positive rate, improve detection delay, and validate this method in different biomes. The method is particularly useful for improving existing forest maps as it is robust to noisy training data. Our results demonstrate that Sentinel-1 data has the potential to advance global deforestation monitoring.


References

Hansen, J. N., Mitchard, E. T., & King, S. (2020). Assessing Forest/Non-Forest Separability Using Sentinel-1 C-Band Synthetic Aperture Radar. Remote Sensing, 12(11).

Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A., Tyukavina, A., Thau, D., Stehman, S. V., Goetz, S. J., Loveland, T. R., Kommareddy, A., Egorov, A., Chini, L., Justice, C. O., & Townshend, J. R. G. (2013). High-Resolution Global Maps of 21st-Century Forest Cover Change. Science, 342(November), 850–854.

Vancutsem, C., Achard, F., Pekel, J. F., Vieilledent, G., Carboni, S., Simonetti, D., Gallego, J., Aragão, L. E., & Nasi, R. (2021). Long-term (1990–2019) monitoring of forest cover changes in the humid tropics. Science Advances, 7(10), 1–22.

Adams, R. P., & MacKay, D. J. C. (2007). Bayesian Online Changepoint Detection.
https://doi.org/arXiv:0710.3742v1

Killick, R., Fearnhead, P., & Eckley, I. A. (2012). Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107(500), 1590–1598.

Recent advancements in satellite-based forest monitoring systems have enabled the detection of forest disturbances in in high spatial detail and in near real-time (e.g. Reiche et al., 2021; Pickens et al., 2020), providing a valuable tool for law enforcement against illegal forest activities. Still, mapping only the location of new disturbances is not always sufficient for timely and targeted intervention. Rapid and automated classification of the direct driver of a new forest disturbance (e.g. agriculture, logging or mining) is a critical next step, that will enable alert prioritization and provide more actionable information. The increasing availability of temporally dense multi-sensor satellite data and developments in artificial intelligence hold an unprecedented potential for implementation of such improvements in forest monitoring methods (Finer et al., 2018). In this study, we focused on the near real-time classification of direct drivers of small-scale tropical forest disturbances, using Sentinel-1 and Sentinel-2 data in combination with convolutional neural networks.

We detected forest disturbances with Sentinel-1 radar data in near real-time at 10m spatial scale, following the methods of RAdar for Detecting Deforestation (RADD) (Reiche et al., 2021). Throughout several study areas in the Congo Basin, Amazon Basin and insular Southeast Asia, reference data was sampled to delineate forest disturbances driven by either smallholder clearing, road development, selective logging or mining. This was done based on visual interpretation of the RADD alerts and high-resolution imagery from quarterly and monthly Planet mosaics (Planet Team, 2017).

We trained a convolutional neural network to classify each newly disturbed forest patch individually, based on a near real-time simulation of RADD forest disturbance detections, starting in 2020 until mid-2021. The neural network was designed to take multi-channel image data as an input, and assign a class label as an output. The input image data consisted of Sentinel-1 and Sentinel-2 imagery from before and after a specific disturbance, the shape and timing of the specific disturbance patch, and the shape and timing of past disturbances in its surrounding, all sampled within bounding boxes of 200x200 10m pixels. In this way, multispectral and backscatter information was captured not only for the disturbed area itself, but also for its direct surrounding. We only considered disturbances no older than 6 months for training, to simulate a near real-time classification scenario.

During the training of our convolutional neural network, the training/validation split was done geographically. Finally, our model testing was done in independent study regions. To refine our classification, we added a class ‘other’ to distinguish forest disturbances that did not fit into any of the defined classes.

We observed macro F1 scores ranging up to 0.78. Most confusion was observed between Road development and Selective logging. This was a somewhat expected outcome, since these two types of forest disturbance are often co-located and have similar visual features in terms of canopy disturbance. Higher macro F1 scores were observed when we distinguished only one driver of forest disturbance (e.g. classifying only mining vs. non-mining resulted in an F1-score of 0.88), indicating that the methods could be applied with higher accuracies for users interested in monitoring only one specific forest disturbance driver, than for distinguishing multiple drivers.

This research presents novel methods for the near real-time classification of direct drivers of forest disturbance in the tropics. Integrating these methods in operational forest monitoring systems will be a major step forward to support timely law enforcement activities in forest management.


References:

- Finer, B. M., Novoa, S., Weisse, M. J., Petersen, R., Mascaro, J., Souto, T., Stearns, F., & Martinez, R. G. (2018). Combating deforestation: From satellite to intervention. In Science. https://doi.org/10.1126/science.aat1203

- Pickens, A. H., Hansen, M. C., Adusei, B., & Potapov, P. (2020). Sentinel-2 Forest Loss Alert. Global Land Analysis and Discovery (GLAD). www.globalforestwatch.org

- Planet Team. (2017). Planet Application Program Interface: In Space for Life on Earth. https://api.planet.com

- Reiche, J., Mullissa, A., Slagter, B., Gou, Y., Tsendbazar, N., Odongo-Braun, C., Vollrath, A., Weisse, M., Stolle, F., Pickens, A., Donchyts, G., Clinton, N., Gorelick, N., & Herold, M. (2021). Forest disturbance alerts for the Congo Basin using Sentinel-1. Environmental Research Letters, 16(2), 024005. https://doi.org/10.1088/1748-9326/abd0a8

Trees Outside Forests (TOF) play a vital role in African ecosystems. In addition to stabilising local climates and acting as a carbon stock, TOF provide key ecological services, and serve as a foundation for local livelihoods through the use of tree products for food, fodder, construction, medicinal use and increased agricultural productivity. Despite their essential role, the extent of TOF has not been assessed at continental scale. Existing tree cover maps primarily quantify forest cover and do not include isolated trees, as these are not discernible in lower resolution satellite images.

Here we use high-resolution images from the PlanetScope nanosatellite constellation to produce a 1 m map of tree cover in 2019 for the African continent. We composite 232053 4-band images into 1x1° mosaics and apply a Convolutional Neural Network to segment canopy cover of all trees and shrubs. To train the network we use a combination of manually annotated 1 m labels, source images upsampled from 3 m to 1 m, and feeding in the output of an initial model trained at lower resolution. This results in a final model that maps tree cover at 1 m across the entire continent, segmenting closed canopies in forest areas and individual scattered trees in savannah areas. We show that for many dryland countries TOF are the dominant component of tree cover, and that most of this tree cover is not considered in existing forest cover maps.

Our analysis demonstrates the new opportunities emerging from the combination of machine learning with commercially-available and low cost nanosatellite imagery, and lays the groundwork towards global scale studies of tree cover at individual tree level and annual temporal scale. This will be crucial for managing woody resources, monitoring TOF in relation to tree planting and restoration efforts, as well as detecting illegal removal of trees.

In recent years, the release of temporally dense and high-resolution remote sensing data from ESA’s Sentinel missions has opened new frontiers for mapping and monitoring changes in global forest cover. However, detecting small-scale degradation in tropical forests remains a challenging task. In tropical and subtropical regions, the availability of optical multispectral imagery is limited by persistent cloud cover, and due to a lack of canopy penetration the time window for detecting change may be too short for an observation to reliably occurred. While synthetic-aperture radar (SAR) data does not suffer from cloud cover issues, poor correlations have been observed between readily-available C-band backscatter and biomass estimates due to signal saturation in high biomass ecosystems, and the same is true for optical vegetation indices. On the other hand, combining information from both types of sensors has shown promising results for estimating vegetation biophysical properties, including forest biomass. Other sources of uncertainties in change detection analysis include the lack of accurate ground data of biomass change and the mismatch between remote sensing acquisition date and field inventory date.

Previous research has suggested that spectral features derived from a single pixel perform poorly at describing canopy structures. However, texture analysis has shown potential for overcoming the existing problems with biomass saturation in dense canopies, both in optical and SAR imagery. Few studies have incorporated the information from both datasets, and rarely including textural information from the optical multispectral sensor. This process has never been tested using experimental measures of biomass change.

To improve detection of small-scale disturbance events in tropical forests, such as those caused by selective logging, we established a degradation experiment, called FODEX and funded by an EU ERC grant, in eight 1-ha plots located in two logging concessions in Peru and Gabon. Biomass data was collected before and after the extraction of a small number of trees, resulting in the loss of between 5 and 34% initial biomass. Spectral and textural information where extracted from Sentinel-1 and Sentinel-2 imagery. Principal component analysis (PCA) was used to remove the noise in the multitemporal images while preserving the indication of change. Multiple linear regression (MLR) methods were used to compute the biomass change estimation models. To overcome the inherent limitations of the two sensors, we build a model incorporating optical and radar textural information and provide optimum window size and textural band combination for detecting fine-scale degradation events.

The results we will present suggest this approach has great promise for mapping forest degradation.

References

Le Toan, T.; Beaudoin, A.; Riom, J.; Guyon, D. Relating forest biomass to SAR data. IEEE Trans. Geosci. Remote Sens. 1992, 30, 403–411

Cutler, M.; Boyd, D.; Foody, G.; Vetrivel, A. Estimating tropical forest biomass with a combination of SAR image texture and Landsat TM data: An assessment of predictions between regions. ISPRS J. Photogramm. Remote Sens. 2012, 70, 66–77.

Luo P, Liao J, Shen G. Combining spectral and texture features for estimating leaf area index and biomass of maize using Sentinel-1/2, and Landsat-8 data. IEEE Access. 2020 Mar 17;8:53614-26.

The recent advancements in remote sensing technology and the ever-growing data availability account for an unprecedented potential for monitoring the Earth's surface. In this context, land cover mapping represents a crucial variable for sustainable decision-making in tasks such as urban expansion planning, natural resources management, and environmental policy development. In this scenario, the Amazon is the world's largest rainforest, characterized by a higher biodiversity than any other ecosystem. Moreover, it comprises the largest river on Earth and the biggest population of indigenous people living in forests. Over the last decades, it witnessed alarming deforestation rates with local and even global repercussions. Thus, a regular and effective monitoring of the Amazon forest is pivotal in the effort to preserve its vital social and ecological roles.

In this context, spaceborne remote sensing stands out as a powerful data source by providing medium- or high-resolution imagery from otherwise inaccessible areas. Within the Amazon scope, there are currently several operational satellite-based systems whose main goal is to map deforestation and land cover changes: the Program to Calculate Deforestation in the Amazon (PRODES), Real-Time System for Detection of Deforestation (DETER), and TerraClass. These projects provide critical information on deforestation, forest degradation, and selective logging that are used for developing effective policies in the context of environmental protection. However, they typically rely on pure visual inspection of optical images over the rainforest, which is laborious, prone to human error, and depends on favorable cloud conditions in a region where the mean annual cloud cover is higher than 70%.

A possible solution for bridging this data gap is synthetic aperture radar (SAR) imaging. SAR systems are microwave sensors that can operate independently of daylight and even under cloudy or adverse weather conditions. The capabilities of using SAR data for forest mapping and large-scale land cover classification have been extensively used in the literature with a great success rate in the last few years, especially with the recent advancements in artificial intelligence (AI) and machine learning (ML), since SAR images are not as easily interpretable with visual inspection as optical data. Nevertheless, most state-of-the-art algorithms for land cover classification with SAR require the use of long time series in order to produce reliable predictions. This situation is not optimal in a fast-changing environment such as the rainforest, where it is of paramount importance to follow the dynamic changes on ground, in order to monitor their status and actively react to illegal deforestation activities.

The Sentinel-1 mission, which comprises a constellation of two polar-orbiting satellites performing C-band SAR imaging, provides an openly available database that has been gaining attention in the field of land cover classification. The image acquisition with its Interferometric Wide swath (IW) mode results in a large swath width combined with a moderate spatial resolution based on the Terrain Observation with Progressive Scans (TOPS) technique. Moreover, the data collection is performed with a revisit time of only 6 days when both satellites are operational and 12 days for a single satellite pass. Thus, Sentinel-1 interferometric SAR (InSAR) short-time-series arise as a potential candidate for systematically monitoring the Amazon rainforest. Apart from the aforementioned reliability of Sentinel-1 imaging, the joint exploitation of SAR backscatter and interferometric coherence and their evolution in time have been shown to be significant features for mapping forests and vegetated areas.

In this framework, recent advances in AI solutions for classification tasks have the potential to optimize feature extraction and pattern recognition with unprecedented accuracy. Deep learning (DL) is a particular AI field devoted to learning complex functions in high-dimensional data with the minimum computational effort possible. For instance, deep convolutional neural networks (CNNs) are responsible for a series of breakthroughs on image classification problems in the last few years and have become an important tool to perform tasks such as image recognition and semantic segmentation in the field of remote sensing.

In this paper, we investigate the potential of CNNs for mapping and monitoring the Amazon rainforest with Sentinel-1 InSAR short-time-series, acquired within a temporal baseline of one month only. In particular, we propose a CNN architecture based on the U-Net, a state-of-the-art method designed for semantic segmentation that combines a powerful data-driven feature extraction with a precise retrieval of spatial information, and train it from scratch to avoid any type of transfer learning. For both training and test, we selected a study site located in the Amazon rainforest covering mostly the Rondônia state, Brazil, in a region that has been suffering from anthropogenic landscape changes over the last decades. As external reference we used the Finer Resolution Observation and Monitoring of Global Land Cover (FROM-GLC) thematic map, whose eight original classes were simplified to map three land cover classes of interest: forests, non-forested regions, and water bodies. Due to the extreme class imbalance towards the forest and non-forest classes, we perform data augmentation on image patches with water bodies to increase their number of samples and statistical variability. Preliminary results of the rainforest mapping show that an overall agreement above 90% can be achieved when compared to the external reference even if only images from a 12-day repeat cycle are used, outperforming the current state-of-the-art shallow learners designed for the same task and suggesting that the proposed approach could be extended to applications at a global scale.

Forest management and conservation require accurate and up-to-date forest maps on tree cover, forest type and tree species composition. This enables forest managers to cope effectively with climate hazards and to take long-term decisions about e.g. thinning operations and species conversions, as well as follow up on the implementation of policy measures.
Flanders, the Northern part of Belgium, is a highly urbanised region in Europe characterised by a highly fragmented landscape. Historical land use and changes have led to severely fragmented forests and a low forest cover percentage of about 10%. Over the past decades, forest multifunctionality has gained importance both in forest policy and practice. Various actions have led to a gradual increase of average stand age and structural diversity. The Flemish government now has the explicit goal to further increase the quantity and quality of forest area in Flanders.
These forests are systematically monitored through ground-based measurement networks, delivering crucial information on the status and long-term evolution of forests. Such a forest inventory, however, only allows for statistical evaluation over longer time periods on the regional scale of Flanders. This translates into a demand for satellite-based approaches that can offer spatially and temporally explicit information on forests.
Since 2017, the Copernicus Sentinel-2A/B satellite constellation records global multispectral remote sensing data with a theoretical revisit time of 2-3 days for mid-latitudes and a spatial resolution varying between 10 and 60 meters for different spectral bands. These data are freely-available through the free, full and open data policy adopted by the Copernicus programme. These characteristics enables the monitoring of vegetation dynamics from space with unprecedented detail both in space and time. Sentinel-2 time series allow for improved characterisation of vegetation phenology and canopy structure.
Meanwhile, traditional machine learning methods are being replaced with deep learning approaches as state-of-the-art in many disciplines after their huge success in computer vision for classification of different data types including digital images and sequential data. These methods, however, have not yet been fully adapted to the specific traits of earth observation data, whose combined spatial, spectral and temporal characteristics are complex. Moreover, training deep learning models typically requires enormous amounts of labelled training samples, which are notoriously lacking in real-life remote sensing applications. At the same time, impressive results have already been obtained for a couple of specific remote sensing tasks, such as land cover classification and crop type mapping. Yet, application in forest remote sensing has been limited.
The core objective of this research is to test the utility of deep learning approaches for Sentinel-2 time series classification in the context of mapping forest type and tree species composition in Flanders using plot-level national forest inventory data. Specific challenges are (1) building a training dataset which copes with the temporal mismatch between reference data collection and Sentinel-2 acquisition dates and (2) training complex models with limited amounts of labelled training samples and unbalanced classes.
In this study, we will employ the Multilayer Perceptron (MLP) as baseline method in which time series data is treated as a multivariate dataset. To exploit the temporal dimension of Sentinel-2 time series, we will implement a 1D Convolutional Neural Network (CNN), namely Temporal CNN (TempCNN)[1], which applies the principle of temporal convolutions.
Labels on forest type and tree species composition were derived from the national forest inventory of Flanders. The spatial sampling design is based on fixed-grid point locations of which approximately 10% lie in forests. Every year, roughly one-tenth of the 2700 plots are visited for dendrological measurements and vegetation measurements. The second measurement cycle was conducted between 2009 and 2019.
This research is performed in the context of the GEO.INFORMED project (https://geo-informed.be). The project’s aim is to develop deep learning workflows that can transform Copernicus Sentinel data into the operational information that is needed by environmental policy agencies. This research was performed in collaboration with the Research Institute of Nature and Forest (INBO). National Forest Inventory data of Flanders is owned, collected and distributed by the Agency of Nature and Forest (ANB).
References
[1] Pelletier, C., Webb, G., & Petitjean, F. (2019). Temporal Convolutional Neural Network for the Classification of Satellite Image Time Series. Remote Sensing, 11(5), 523. https://doi.org/10.3390/rs11050523